caph-dev/agents-progressive-disclosure

agents-progressive-disclosure is a skill for AI coding agents that helps clean up and restructure bloated agent instruction files. These are files like AGENTS.md or CLAUDE.md that tell an AI assistant how to behave in a project. Over time, teams keep adding more and more rules to these files until they become large, dense documents that the agent has to load in full on every single task, even when most of the content is not relevant to the work at hand. The skill addresses two problems this creates. First, the most important rules get buried among rarely-needed details and lose their prominence. Second, every task pays a context cost for instructions it does not need, which wastes the limited space available for the actual work. The solution the skill applies is called progressive disclosure. The idea is to turn the root instruction file into a short routing document that lists core always-on rules and points to separate focused files for task-specific details. Instead of one large file covering everything, you end up with a compact entrypoint that tells the agent where to look for deeper guidance when a specific situation comes up. The workflow the skill follows involves reading the existing instruction file in full, scanning it for contradictory rules before moving anything, classifying each rule as either always-on or task-specific, designing a map of separate reference files, rewriting the root file as a compact router, and then validating that no rules were lost in the process. Installation works through the skills CLI with a single command, or by asking the agent to install it directly from the repository URL. It supports multiple agent environments including Cursor, Codex, Claude, and others. The repository itself is small: it contains the skill definition file, a metadata file, and the README. A full Chinese translation of the README is also included.